# Learning-loop review — 2026-08-28

385 closed trades reviewed, 189 currently open. Proposal only — nothing here is applied automatically (CLAUDE.md §4: the AI review proposes, dan disposes).

## 1. Is overall expectancy holding?

Pooled expectancy is **+1.73% per trade** over **385 closed trades**, PF 2.24. The trade count clears the ~20-trade minimum by a wide margin, so at the "all trades" level this is a real, non-noise signal — expectancy is holding and looks healthy in aggregate.

Caveat worth stating explicitly: this pools three accounts with **different risk configs** (paper-main, Paper02qsr, live-1), and QSR alone contributes 289/385 trades and $4,210 of the $4,268 total P&L. So "overall expectancy" is really "expectancy dominated by QSR under paper-main's config" — not an independent confirmation from three separate experiments.

## 2. Which bucket is the biggest drag?

Restricting to buckets with ≥20 trades, **orb-v1** (n=68) is the standout drag: expectancy **-0.00%**, profit factor **0.92** (<1, i.e. losing money), win rate 45.6% with Wilson lower bound **34.3%**, and net P&L **-$67**. This clears the 20-trade minimum.

Gap vs. the "rest": overall expectancy is 1.73%; orb-v1's is ~0%, a gap of ~1.73pp. That is **smaller than the 2.48pp noise floor** given for bucket-vs-bucket comparisons. So even though orb-v1 clears the trade-count bar, the *size of its underperformance relative to other strategies* is not resolvable with confidence — it could be a real weak strategy or just be within the noise band. (Separately, its PF<1 and net-negative dollar P&L stand on their own regardless of comparison to "the rest," which is a different, more defensible observation than the relative-gap claim.)

## 3. Entry, exit, or regime problem?

The report only gives MAE/MFE at the **overall pooled** level (MAE winners -1.73%, MAE losers -4.32%, MFE losers +0.83%), not broken out per strategy — so a rigorous entry-vs-exit diagnosis specific to orb-v1 isn't available here.

Working with what we have: the pooled MFE-losers figure (+0.83%) is small, not "materially positive" — it doesn't show a strong signature of "trade worked, then gave it back," which would point to an exit problem. Orb-v1's own numbers (45.6% win rate, Wilson low 34.3%, 0.3-day average hold) look more consistent with an **entry-quality problem** — many quick failed breakouts rather than winners mismanaged on the way out. This read is tentative given the lack of orb-v1-specific MAE/MFE data.

No evidence points to a regime problem specifically for orb-v1 — regime is only broken out at the "all trades" level (SPY>MA200 vs unknown), not per strategy.

## 4. Proposed parameter change

Given point 2's finding — orb-v1 clears the 20-trade bar but its gap vs. other strategies does **not** clear the noise floor — any change to an entry-signal parameter (`volumeConfirmMult`, `entryCutoffMinute`, `targetRMultiple`, etc.) chosen to "fix" orb-v1's win rate would be exactly the kind of threshold-tuning-on-noise the rules warn against. **NO entry-signal change is proposed.**

Instead, the one change I'd propose is a **risk-sizing** adjustment, which doesn't depend on out-predicting the market and is justified purely by orb-v1's absolute record (PF 0.92 < 1, net -$67 over 68 trades) rather than by a comparison that needs statistical resolution:

- **Parameter:** `orb-v1.sizeMultiplier`
- **Current:** `0.5`
- **Proposed:** `0.25`
- **Reasoning:** orb-v1 is already the smallest-sized strategy (0.5x), and at n=68 it's the only strategy bucket with PF<1 and flat/negative expectancy on its own terms (not by comparison to peers). Halving size again reduces capital exposure to a currently-unproven strategy without touching any entry/exit logic, preserves the ability to keep collecting data, and is trivially reversible. (Note: orb-v1 is listed as backtest-only/not live-wired currently, so this change would take effect only if/when it's wired live — flagging that context for whoever applies it.)

## 5. Notable open positions

- Several **qsr** positions are aging a long time with `maxHoldDays: null` (no calendar time-stop) and sitting underwater: **HDB** (37.3d, -4.30%), **ARM** (9.0d/1.2d legs, -6.97%), **HON** (16.3d, -5.54%), **STM** (9.3d, -6.60%), **ETR** (23–28d, -2.03%), **TRP/BTI/CVS/PWR/URI/ENB** all red and open 8–28 days. With no time-stop, these can run indefinitely — worth watching given the closed-trade evidence shows `stop_loss` exits are almost always losers (win rate 1.0%) once they do trigger.
- The three **crypto-trend-v1** positions (BTC, ETH, DOGE) have been open 8–38 days with only 1 closed crypto trade in the whole journal — essentially unproven, no time-stop either.
- A large cluster of very recent (1–2 day) small QSR positions (PDD, CARR, NKE, CDNS, URI, HWM, META) suggests a burst of fresh entries — too new to read anything into yet.
- This is color only — none of it feeds into point 4's proposal, since these trades have no confirmed outcome.

## 6. QSR shadow comparison note

The shadow-log window (2026-08-09 → 2026-08-23) is **closed**, so a fuller outcome backtest comparing legacy vs. newlyA hypothetical/real results can now be requested.

Qualitatively: the two selection methods diverge substantially. Legacy (isTriggered+isBuy) fired only 48 times across 18 tickers in the window; the live newlyA method has since produced 120 closed + 170 open entries across 35 tickers — a much higher firing rate and broader universe. Ticker overlap is partial (10 of legacy's 18 tickers — AEP, SU, COHR, TRP, BTI, ENB, NKE, ITUB, EBAY, VALE — also appear under newlyA), meaning roughly 8 legacy-favored tickers (e.g. MO, IBKR, DDOG, AZN, TDG, MAR, MRVL, SNDK) are essentially untouched by the live method, while newlyA also trades ~25 tickers legacy never would have flagged. This is a meaningful behavioral difference in signal timing/breadth worth a dedicated backtest, but per the instructions it is observational only — none of this feeds into point 4, since the shadow trades never executed.